A Machine Learning approach for predicting optimal aerodynamic parameters obtained from fluid dynamic simulations

This work presents an innovative methodology that integrates machine learning into the aerodynamic optimization, focusing on multi-objective optimization and constrained scenarios. This work explores the application of machine learning as an enhancement of the optimization algorithm. Gradient Boosting and Random Forest were primarily used, demonstrating their efficacy in competing with the existing pilOPT algorithm during the optimization process. Gradient Boosting, in particular, emerged as the superior machine learning technique due to its lower mean squared error and more refined design space exploration. The proposed methodology operates in a three-phase approach: an initialization phase, an iteration phase followed by a fine-tuning phase, where the transition over these phases operates at different aspects for exploration and exploitation. The constrained optimization tests yielded mixed results as in synthetic scenarios involving a lower number of input variables, machine learning effectively was able identify designs into the defined constraints, while in a real scenarios with higher input dimensionality, results were suboptimal due to the model limited generalization capabilities and the constraints being defined over a very small subset of the Pareto front. Despite this challenges, the integration of machine learning has been shown to provide advantages, particularly in early iterations where it competes well against the pilOPT algorithm. This work highlights areas for further enhancement, including refining the Evaluation table, improving machine learning algorithms, and potentially incorporating physics-informed methodology for CFD simulations. These potential improvements could lead to significant advancements in both optimization accuracy and computation time.